核心概念
Sparse attention can significantly improve efficiency without compromising effectiveness in cross-encoders.
要約
The content explores the impact of sparse attention on cross-encoders, focusing on reducing token interactions for efficiency while maintaining re-ranking effectiveness. It covers experiments, related work, theoretical explanations, and empirical evaluations.
Directory:
Abstract
Sparse attention enhances efficiency in cross-encoders.
Introduction
Pre-trained transformer models are crucial for retrieval systems.
Related Work
Comparison between bi-encoders and cross-encoders.
Sparse Attention Mechanisms
Windowed self-attention and cross-attention patterns.
Experimental Setup
Fine-tuning models with different window sizes.
Empirical Evaluation
Effectiveness results on TREC Deep Learning tasks.
Out-of-domain Effectiveness
Comparison with other cross-encoder models on TIREx benchmark.
Efficiency Results
Comparison of efficiency metrics for different models.
統計
Our code is publicly available.
引用
"Cross-encoders allow queries and documents to exchange information via symmetric attention."
"Our proposed sparse asymmetric attention pattern combines windowed self-attention and token-specific cross-attention."